A polynomial-time heuristic for the quay crane double-cycling problem with internal-reshuffling operations

A polynomial-time heuristic for the quay crane double-cycling problem with internal-reshuffling operations

Transportation Research Part E 81 (2015) 52–74 Contents lists available at ScienceDirect Transportation Research Part E journal homepage: www.elsevi...

2MB Sizes 2 Downloads 92 Views

Transportation Research Part E 81 (2015) 52–74

Contents lists available at ScienceDirect

Transportation Research Part E journal homepage: www.elsevier.com/locate/tre

A polynomial-time heuristic for the quay crane double-cycling problem with internal-reshuffling operations Ming Liu a, Feng Chu b, Zizhen Zhang c,⇑, Chengbin Chu a,d a

School of Economics & Management, Tongji University, Shanghai, PR China Laboratory IBISC, University of Evry-Val dÉssonne, Evry 91020, France c School of Mobile Information Engineering, Sun Yat-sen University, Gangzhou 510275, PR China d Laboratoire Génie Industriel, Ecole Centrale Paris, Grande Voie des Vignes, 92295 Châtenay-Malabry Cedex, France b

a r t i c l e

i n f o

Article history: Received 19 November 2014 Received in revised form 24 June 2015 Accepted 25 June 2015 Available online 10 July 2015 Keywords: Container port Terminal operations Double cycling Internal reshuffling

a b s t r a c t One of great challenges in seaport management is how to handle containers under reshuffling, called reshuffles. Repositioning reshuffles in a bay (internal reshuffling) can improve the efficiency of quay cranes and help ports to reduce ship turn-around time. This paper studies the quay crane double-cycling problem with internal-reshuffling operations, and presents a fast solution algorithm. To reduce the number of operations necessary to turn around a bay of a vessel, the problem is first formulated as a new integer program. A polynomial-time heuristic is then developed. The analysis is made on the worst-case error bound of the proposed algorithm. Results are presented for a suite of combinations of problem instances with different bay sizes and workload scenarios. Comparisons are made between our algorithm and the start-of-the-art heuristic. The computational results demonstrate that our model can be solved more efficiently with CPLEX than the model proposed by Meisel and Wichmann (2010), and the proposed algorithm can well solve real-world problem instances within several seconds. Ó 2015 Elsevier Ltd. All rights reserved.

1. Introduction Efficient seaports could help to lower transport costs by enabling cargos to get to and from markets in a more timely and cost-effective fashion. Nowadays, more and more goods in global trade are containerized and transported by container vessels. High utilization of container terminal resources are key in accelerating maritime logistics and lower operating costs. Quay cranes (QCs) are the most expensive single unit of handling equipment in container ports. One of the key operational bottlenecks at ports is QC availability (Crainic and Kim, 2005). To reduce ship turn-around time, ports make continuous improvement on the efficiency of QCs. As QC efficiency is the key bottleneck to port productivity, the work presented in this paper addresses such an operational-research problem. In contrast to terminal expansion or information technology deployments, the internal-reshuffling method, considered in this paper, is low-cost. The internal-reshuffling technique can be quickly implemented, and can be used to complement the classic double-cycling technique (i.e., loading ships as they are unloaded). The layout of containers on a ship can be viewed as a box. Containers are stacked on top of one another and arranged in rows, also called bays. Fig. 1 illustrates the plan, side and front views of a container vessel: in the plan view, there are seven ⇑ Corresponding author. E-mail address: [email protected] (Z. Zhang). http://dx.doi.org/10.1016/j.tre.2015.06.009 1366-5545/Ó 2015 Elsevier Ltd. All rights reserved.

M. Liu et al. / Transportation Research Part E 81 (2015) 52–74

53

Fig. 1. Three views of a vessel (Goodchild and Daganzo, 2006).

rows where the first row consists of six container stacks; in the side view, a stack holds six containers; in the front view, six container stacks are listed from left to right on the vessel. (In Fig. 1, the number of containers are not representative of typical ship size.) Note that containers can be accessed only from above by QCs, and such a stacking manner induces the precedence constraints among containers of each stack, also called the stack dependent accessibility. Considering that QCs move slowly, only after completing all operations in one bay a crane operator may drive it to the next one, as is current practice. Traditionally, the unloading and loading processes of a vessel are separated. In one cycle of quay crane’s trolley, it unloads (or loads) one container and return without container, referred as single cycling (see Fig. 2a). To improve QC productivity, double-cycling technique is first proposed in a pioneering work (Goodchild, 2005). The technique consists of converting empty crane moves into productive ones. With double cycling (see Fig. 2b), containers are loaded and unloaded in the same crane cycle, i.e., a complete round-trip of the crane trolley from the ship to the dock and back. This allows a QC to double the number of containers transported in one cycle (Goodchild and Daganzo, 2007). Nowadays, the double-cycling technique has been widely adopted and implemented, e.g., at train terminals (Goodchild et al., 2011). Although double cycling has been used to some extent in practice, port operators ask for more QC efficiency to counter port congestion which has been hitting the headlines recently (Drewry Container Insight, 2014). To improve QC efficiency, internal reshuffling, a complementing technique to the double-cycling technique, can be used to improve the efficiency of QCs by replacing couples of unloading and loading operations for reshuffles by repositioning operations (Meisel and Wichmann, 2010). Instead of using the current method (called external reshuffling), where often all reshuffles are unloaded from the vessel and then reloaded from the dock, some reshuffles can be repositioned directly in a bay on a ship. This allows the QC to replace two operations (i.e., one unloading and one loading operation) by one operation (i.e., a repositioning), to complete a reshuffling requirement. Thus, this technique reduces the number of operations, and ensures high utilization of QCs. In the remainder, for easy recognition where internal-reshuffling method is applied, we also refer to a repositioning operation induced by internal-reshuffling method as an internal-reshuffling operation. We assume the

Fig. 2. Single cycling and double cycling (Goodchild and Daganzo, 2006).

54

M. Liu et al. / Transportation Research Part E 81 (2015) 52–74

ship’s arrival stowage plan, also called arrival plan, and departure stowage plan, also called departure plan, are given, which indicate the positions, also called slots, of the containers to be unloaded, loaded and reshuffled. In current practice, shipping lines use software tools to create such plans that accommodate vessel stability requirements, priority of delivery, placement constraints on hazardous materials, etc. (Goodchild and Daganzo, 2006). We therefore consider improvements only on the crane’s sequence of operations, to complete the conversion from the arrival plan to the departure plan. Note that not all reshuffles can be handled internally. Only if a slot to be hold by a reshuffle in the departure plan is empty, may an internal-reshuffling operation happen. In the classic quay crane double-cycling problem (QCDCP), decisions are made on the basis of container stacks. That is, a solution consists of a sequence of container stacks. In our problem, the quay crane double-cycling problem with internal-reshuffling operations (QCDCP-IR), a solution must be made on a container basis (i.e., it consists of a sequence of container operations). The QCDCP-IR is more complicated and requires more sophisticated solution algorithms. Meisel and Wichmann (2010) initiate the study of the QCDCP-IR, and propose an integer programming formulation and a heuristic. In their work, the advantage of the internal-reshuffling method in improving QC productivity has been shown, compared with the sole application of the double-cycling technique. Their experimental results demonstrate that about 64% of the reshuffles in a bay can be reshuffled internally by their solution approach, and 32% operations related to reshuffles can be saved by the internal-reshuffling method. However, their mathematical formulation is relatively rough, and their heuristic approach has no guaranteed error bounds and requires a relatively large computational time, especially for practical-size problems. Thus, a fast solution approach with acceptable worst-case error bounds should further improve QC efficiency for the QCDCP-IR. We revisit the QCDCP-IR, aiming to provide some analytical results and an easy-to-implement solution approach. The contribution includes  We formulate a new integer program based on new several observations. Comparisons are made between the existing model and our new formulation, both solved with CPLEX. Experimental results show that our model is so-called CPLEX-effective, which means (i) delivering the same quality solutions, our formulation saves computational time and (ii) consuming the same running time, our formulation delivers higher quality solutions.  To address the general QCDCP-IR, especially for large-size problems, we devise a fast heuristic, and analyze its worst-case error bound. Computational experiments based on 1000 instances have been conducted. Compared with the state-of-the-art heuristic, results show that our solution approach is very efficient in terms of running time and it is also relatively robust. The paper is organized as follows. A brief literature review is given in Section 2. In Section 3, we restate the QCDCP-IR. In Section 4, we provide a new integer programming formulation based on several observations. In Section 5, for the general case, we devise a polynomial-time heuristic and analyze its worst-case error bound. In Section 6, computational experiments are conducted. The performance of our model and the existing one are first compared. Then we evaluate the solution quality of the proposed algorithm by comparison with the state-of-the-art heuristic, to demonstrate the efficiency of our approach. Section 7 concludes the paper and give directions of further work. In Appendix, we correct errors in Meisel and Wichmann (2010)’s formulation, and identify a polynomially solvable case for the QCDCP-IR.

2. Literature review In the literature, a comprehensive overview on applications and optimization models for container port optimization can be found in Stahlbock and Voß (2008), Bierwirth and Meisel (2010), and Carlo et al. (2013). Other study streams of maritime logistics may include liner shipping network management (Meng and Wang, 2011; Wang, 2014 and Wang et al., 2014), container assignment and routing (Bell et al., 2011; Bell et al., 2013 and Wang et al., 2015), and vessel bunker and speed management (Wang and Meng, 2012 and Wang et al., 2013). Recent advances on liner shipping and port operations planning can be found in Meng et al. (2014) and Bierwirth and Meisel (2015), respectively. As our concern falls in the scope of the quay crane scheduling at seaports, we review the most pertinent literature. Daganzo (1989) analyzes the performance of different quay crane scheduling algorithms, with the objective to maximize the throughput. Kim and Park (2004) define the tasks on the basis of container groups, which are subsets of containers on a vessel with identical destinations or other attributes such as weight class. An exact algorithm and a greedy randomized adaptive search procedure are proposed. Lim et al. (2007) analyze a relatively simple model with a major feature of the non-crossing restriction. They prove the existence of an optimal schedule which is unidirectional if the bay-to-crane assignment is given. This rule has been widely adopted especially for the development of heuristics. Bierwirth and Meisel (2009) revise Kim and Park (2004)’s model and propose a heuristic based on a branch-and-bound framework. Choo et al. (2010) investigate a multi-ship quay crane scheduling problem with yard congestion constraint. They develop a Lagrangian relaxation-based heuristic. Recently, Liu et al. (2014) design a 4/3-approximation and 5/3-approximation algorithms for the quay crane scheduling problems with two and three QCs, respectively. Goodchild and Daganzo (2006) initiate the study of the QCDCP. They formulate the unloading and loading operations of each container stack as a two-machine flow shop job with two sequential processing times. They show that when there are no hatch covers, the problem can be mapped into a two machine flow shop scheduling problem, and be solved optimally

M. Liu et al. / Transportation Research Part E 81 (2015) 52–74

55

with Johnson’s rule. They further study a general QCDCP in which hatch covers are involved, and propose a decomposition heuristic. Containers under reshuffling are handled externally in this work. For the general QCDCP, Zhang and Kim (2009) develop a local-search based heuristic. Numerical experiments show their solution approach is very efficient. Recently, Lee et al. (2014) have developed an optimal algorithm for the general QCDCP, which runs in polynomial time. Meisel and Wichmann (2010) initiate the study of the QCDCP-IR, where internal-reshuffling operations are enabled. The problem is investigated on a container basis. An integer programming formulation and a heuristic approach are provided. They also demonstrate that the consideration of the internal-reshuffling operations leads to shortening of the vessel handling time compared to the sole application of the double-cycling technique. 3. Problem description In this section, we restate the QCDCP-IR, which involves a single bay and a single quay crane. In one crane cycle, the QC is enabled to perform at most one unloading and one loading operations. Given the arrival and departure plans, the aim is to determine a feasible sequence of container operations that converts the arrival plan to the departure plan with minimum service time. A feasible solution must respect the stacking dependent accessibility. In other words, the precedence constraints induced by the container stacking manner cannot be violated. Containers are categorized into four classes: import containers, export containers, fixed containers, and reshuffles (i.e., containers to be reshuffled). Import containers are those to be unloaded from the vessel to the dock, whereas export containers are those to be loaded from the dock to the vessel. Clearly, import (export) containers appear only in the arrival (departure) plan. There are also some containers on the ship not destined to the current terminal, which can be classified as (i) reshuffles which stay on top of the import containers, must be temporarily removed and then restored, as they block the unloading operations of the import containers and (ii) fixed containers which are to stay on the vessel involving no operations, and thus are not considered in the following analysis. Notice that the same reshuffles can be found in both plans. From the perspective of terminal managers, reshuffles belong to the same class and thus they can exchange positions in the departure plan (Meisel and Wichmann, 2010). That is, each reshuffle in the arrival plan can be placed at any slot to be hold by a reshuffle in the departure plan. Reshuffles can be handled externally or internally. Apparently, an internal-reshuffling operation is preferable to an unloading operation plus a loading operation, as the former requires only one container operation, whereas the latter needs two. However, only when a slot to be hold by a reshuffle in the departure plan (thus it is ready to receive a reshuffle) is available, may an internal-reshuffling operation occur. Fig. 3 illustrates an example. As there is a doubt how a change of the sequence of operations brings benefits to reduce ship turn-around time, we compare a solution without internal-reshuffling operations and a solution where internal-reshuffling technique is implemented. In the former solution, the sequence must consist of 15 operations, i.e., nine unloading and six loading operations, where reshuffles are handled externally. Fig. 4 illustrates the latter solution with internal-reshuffling operations. This solution comprises six unloading, five loading and two internal-reshuffling operations, thus 13 operations in total. For the latter solution, a step-by-step illustration of service time will be shown in Fig. 10. Note that at steps 5 and 10, internal-reshuffling operations happen. Compared with the former solution of 15 operations, the internal-reshuffling method helps to eliminate two container operations, completing the operations necessary to turn around a bay of a ship. 3.1. Assumptions For the QCDCP-IR, the following assumptions are made (Meisel and Wichmann, 2010). A1. Each of import and export containers is processed by one operation. For example, this prevents the repositioning of an import container to another slot in the bay before it is finally unloaded. A reshuffle handled internally needs one operation, whereas a reshuffle handled externally requires two operations.

Fig. 3. The arrival and departure plans (Meisel and Wichmann, 2010).

56

M. Liu et al. / Transportation Research Part E 81 (2015) 52–74

Fig. 4. A sequence of container operations, including six unloading, five loading and two internal-reshuffling operations.

A2. Reshuffles are exchangeable, i.e., every reshuffle in the arrival plan can be positioned at any slot to be hold by a reshuffle in the departure plan. The number of reshuffles in one stack is the same in both plans. This ensures that after the unloading operations of all the containers in a stack, a sufficient number of reshuffles is available for the subsequently loading operations for the stack. A3. The stability issue of the vessel is not considered. The arrival and departure plans are preset for the planning of quay crane operations. A4. A hatch-coverless container vessel is considered, where containers below and above deck are not separated by hatch covers. A5. Horizontal transport vehicles are always available at the dock, waiting to receive unloaded containers or provide containers to be loaded. Hence, the service time of the bay only depends on the sequence of container operations. A6. The quay crane starts processing of containers from the vessel. In fact, no matter where the quay crane starts, at the vessel or dock, it completes the same work.

Remark 1. Assumption A2 is relatively simplistic, because in reality a reshuffle may be moved from one bay to another. This happens, for instance, to consolidate containers with the same destination, which were not consolidated in the same bay due to previous port operations. This assumption is to make the QCDCP-IR restricted in a single bay, as the classic QCDCP. 3.2. Notation Let m denote the number of container stacks, and n the total number of containers in both plans. Five types of jobs (corresponding to container operations) are identified: (i) VY: unload an import container from the vessel to the yard; (ii) YV: load an export container from the yard to the vessel; (iii) VB: unload a reshuffle from the vessel to the buffer on the yard; (iv) BV: reload a reshuffle from the buffer to the vessel; and (v) VV: reposition a reshuffle from one slot on the vessel to another slot on the vessel. With these notation, unloading operations consist of VY and VB jobs, whereas loading operations consist of BV and YV jobs. Internal-reshuffling operations correspond to VV jobs. Let T denote the set of all the job types, i.e., T = {VY,YV,VB,BV,VV}. Let pt denote the processing time of the t-type job, where t 2 T , and stu the setup time (corresponding to the required time of empty trolley move) between the t-type job and the u-type job ðt; u 2 T Þ, such that the latter immediately follows the former in a sequence. (In the scheduling literature, setup time means a time period required by the machine to be ready to process a job.) VV job has a less processing time than those of other types, i.e., pVV ¼ mint2T pt , because it starts and ends on the vessel, whereas other jobs travel between the vessel and the dock. The following Tables 1 and 2 are presented in Meisel and Wichmann (2010), where the time unit is second. In Table 1, a VV job has a processing time of 90 s whereas each of the others has a processing time of 100 s. In Table 2, if the two jobs are connected on the vessel or the dock, then the setup time in between requires 10 s, otherwise 20 s.

57

M. Liu et al. / Transportation Research Part E 81 (2015) 52–74 Table 1 Processing time pt . t

pt

VY YV VB BV VV

100 100 100 100 90

Table 2 Setup time stu . t # nu !

VY

YV

VB

BV

VV

VY YV VB BV VV

20 10 20 10 10

10 20 10 20 20

20 10 20 10 10

10 20 10 20 20

20 10 20 10 10

4. New mathematical formulation In this section, we first transform the QCDCP-IR to an equivalent scheduling problem, and then formulate a new integer program. We follow Goodchild and Daganzo (2006)’s work by extending their scheduling approach to the QCDCP-IR and thus anticipate that the QCDCP-IR can be viewed as a single machine makespan minimization scheduling problem (with precedence constraints and sequence-dependent setup time) and the service time of a sequence of jobs in the QCDCP-IR is equal to the makespan of a schedule in the scheduling problem. For simplicity, we use a concise representation of the problem hereafter. For ease of illustration, we index all the containers in both plans. Fig. 5 illustrates a new representation of the example given in Fig. 3 with indices, where n, the number of all containers, is equal to 15, which are stowed in m ¼ 4 stacks. In Fig. 5, all fixed containers are dropped, as they are not in our consideration. 4.1. Problem transformation A feasible solution to the QCDCP-IR must respect the stacking dependent accessibility. Fig. 6 gives the precedence digraph of the example problem. Each container in both plans is represented by a node. Because the precedence constraints only exist among containers in each stack (e.g., job 1 must be processed before job 2, and job 11 must be processed before job 12), and the containers in the arrival plan must be processed before those in the departure plan (e.g., job 2 must be processed before job 10), and thus the precedence relation of containers in each stack can be mapped as a chain. Therefore, in the precedence digraph, there are m parallel chains (here m ¼ 4) for all n jobs (here n ¼ 15). Each node is also associated with a two-tuple ði; jÞ, which indicates that the container is located at the j-th position (from the chain head) in the i-th chain in the digraph. Such a set of notation is useful in our new formulation, which we will explain later. In Fig. 6, redundant arcs are reduced by the transitive property of precedence relations. To make internal-reshuffling operations (corresponding VV jobs) easy to handle, we reduce the solution search space by the following observations.

Fig. 5. A concise representation of the example.

58

M. Liu et al. / Transportation Research Part E 81 (2015) 52–74

Fig. 6. Precedence digraph of the example problem.

Observation 1. A VV job is equivalent to a VB job immediately followed a BV job, with a virtual setup time of s ¼ pVV  ðpVB þ pBV Þ in between (e.g., see Fig. 7). For any sequence which contains VVs, we can replace these VV jobs according to Observation 1, and thus no sequences under our consideration contain VVs hereafter. As compensation, for such sequences, we need to further record where each VV is replaced by a couple of VB and BV and add a setup time s in between. This observation serves as a preprocessing step, and after this step each sequence in the search space consists of exact n jobs (i.e., the number of containers in both plans). Now there exists a phenomenon that in some sequences, the setup time between consecutive VB and BV is sVB;BV (e.g., 10 s as in Table 2), and in some sequences that value is s (e.g., 110 s according to Table 2). We have sVB;BV > s, because the former corresponds to an unloading operation immediately followed by a loading operation for a reshuffle and the latter corresponds to an internal-reshuffling operation. We have the following observation. Observation 2. In a job sequence, if a VB job is immediately followed by a BV job with setup time sVB;BV in between, then this sequence is dominated by a corresponding sequence where each setup time sVB;BV is replaced by a setup time s. Applying the above two observations, we efficiently reduce the search space. Now Tables 1 and 2 can be updated to Tables 3 and 4. Note that in these tables, VVs are dropped (by Observation 1), and the setup time between VB and BV is replaced by ‘‘110’’ (by Observation 2). Moreover, we have the following one-to-one relation, which maps each job (representing a container operation) to a container in both plans. Observation 3. In a job sequence, (i) each VY job corresponds to an import container, (ii) each VB job corresponds to a reshuffle in the arrival plan, (iii) each YV job corresponds to an export container, and (iv) each BV job corresponds to a reshuffle in the departure plan. By this observation, we can use VY, VB, YV and BV to denote the corresponding containers, respectively, and thus each container is also referred as a job with a certain type. (A sequence of containers can be used to portray a sequence of

Fig. 7. A sequence with VV and a sequence without VV.

59

M. Liu et al. / Transportation Research Part E 81 (2015) 52–74 Table 3 Processing time pt . t

pt

VY YV VB BV

100 100 100 100

Table 4 Setup time stu . stu

VY

YV

VB

BV

VY YV VB BV

20 10 20 10

10 20 10 20

20 10 20 10

10 20 110 20

container operations). The QC service time is the total time necessary to complete all the n jobs. The QCDCP-IR can be viewed as a kind of single machine makespan minimization scheduling problem, with precedence constraints and sequence-dependent setup time, which is NP-hard. Specifically, the quay crane is regarded as the single machine. Before the first job in a sequence, there is no setup time, or to say the setup time is zero. The scheduling problem aims to determine a sequence of jobs with minimum makespan (i.e., completion time). 4.2. Integer programming formulation Based on the precedence digraph, we formulate a new integer program for the QCDCP-IR. Given is a precedence digraph with m chains and n jobs (e.g, see Fig. 6). Let J = {VY, VB, YV, BV} denote the set of job types. We also use the following notation. Indices: ði; jÞ: t; u: k:

job index, indicating the job located at the j position in the i-th chain, where positions are indexed in a chain from head (top) to tail (bottom); job type indices, where t; u 2 J ; position index in the sequence, k 2 f1; 2; . . . ; ng.

Parameters: VY: set of VY-type jobs; YV: set of YV-type jobs; VB: set of VB-type jobs; BV: set of BV-type jobs; S S S G: set of jobs of all types, i.e., G ¼ VY YV VB BV; pt : processing time of a t-type job, where t 2 J ; stu : setup time of a u-type job, immediately preceded by a t-type job, where ft; ug 2 J . For the digraph in Fig. 6, we know VY ¼ fð1; 2Þ; ð2; 1Þ; ð3; 1Þ; ð3; 3Þ; ð4; 3Þg, VB ¼ fð1; 1Þ; ð3; 2Þ; ð4; 1Þ; ð4; 2Þg and BV ¼ fð1; 3Þ; ð3; 5Þ; ð4; 4Þ; ð4; 5Þg.

YV ¼ fð3; 4Þ; ð4; 6Þg,

Decision variables: xijk : equal to 1 if job ði; jÞ is arranged at the sequence’s k-th position (i.e., the k-th processed); 0 otherwise. stk : equal to 1 if the sequence’s k-th position is occupied by a t-type job, t 2 J ; 0 otherwise. etu k : equal to 1 if the sequence’s k-th position is possessed by a t-type job, and the ðk þ 1Þ-th position is occupied by a t u u-type job, where k 2 f1; 2; . . . ; n  1g; 0 otherwise. That is, etu k ¼ sk  skþ1 . Let M1 denote a fixed integer, larger than all stu . In our formulation, we use a modified objective function Z 0 instead of the original objective function value Z, such that Z ¼ Z 0  M1ðn  1Þ. For the minimization problem, the usage of Z 0 is to make tu variable etu k preferable to be zero, guaranteed by positive multipliers fs þ M1g. The graph-based integer program (GBIP) is given below.

60

M. Liu et al. / Transportation Research Part E 81 (2015) 52–74

min Z 0 ¼

ðGBIP modelÞ

n X n1 X X X pt  stk þ ðstu þ M1Þ  etu k k¼1 t2J

n X xijk ¼ 1;

s:t:

ð1Þ

k¼1 ft;ug2J

ði; jÞ 2 G:

ð2Þ

k¼1 n n X X xijk  k þ 1 6 xi;jþ1;k  k; k¼1

fði; jÞ; ði; j þ 1Þg 2 G:

ð3Þ

k¼1

X

stk ¼ 1; k 2 f1; 2; . . . ; ng:

ð4Þ

t2J

sVY k ¼

X

xijk ;

k 2 f1; 2; . . . ; ng:

ð5Þ

xijk ;

k 2 f1; 2; . . . ; ng:

ð6Þ

xijk ;

k 2 f1; 2; . . . ; ng:

ð7Þ

xijk ;

k 2 f1; 2; . . . ; ng:

ð8Þ

ði;jÞ2VY

sYV k ¼

X ði;jÞ2YV

sVB k ¼

X ði;jÞ2VB

sBV k ¼

X ði;jÞ2BV

t u etu k P sk þ skþ1  1;

xijk 2 f0; 1g;

ft; ug 2 J ; k 2 f1; 2; . . . ; n  1g:

ði; jÞ 2 G; k 2 f1; 2; . . . ; ng:

ð9Þ ð10Þ

stk 2 f0; 1g; t 2 J ; k 2 f1; 2; . . . ; ng:

ð11Þ

etu k 2 f0; 1g;

ð12Þ

ft; ug 2 J ; k 2 f1; 2; . . . ; ng:

The objective function (1) minimizes the makespan (i.e., completion time), which is the sum of job processing times and setup times. Note that the original objective value is restored by adding M1ðn  1Þ to Z 0 . Constraints (2) ensure that each job ði; jÞ is processed exactly once, by summing up all positions k for each job. Constraints (3) guarantee that precedence relaPn tions are satisfied. In particular, item k¼1 xijk  k denotes the position in the sequence occupied by job ði; jÞ and item Pn k¼1 xi;jþ1;k  k denotes the one possessed by job ði; j þ 1Þ. As job ði; jÞ must be processed before job ði; j þ 1Þ due to precedence relations, the position of ði; jÞ in the sequence must be earlier than the position of job ði; j þ 1Þ by at least one unit. Constraints (4) impose that only one job (of any type, by summing up all types) can be processed at each position k in the sequence. Constraints (5)–(8) guarantee the corresponding relation between stk and t-type jobs. Constraints (9) state the relation between the setup time and the related two jobs. Note that etu k is minimized by the objective function. Domains of the decision variables are defined by (10)–(12). Our GBIP model is suitable for different values of processing times and setup times (i.e., not restrictive to the data in t u tu t tu u Tables 3 and 4). To express the relation etu k ¼ sk  skþ1 , two sets of constraints ek 6 sk and ek 6 skþ1 are theoretically required. to be zero, these two set of constraints are redundant in our formulation. (This However, as the objective function Z 0 forces etu k may be one reason why our formulation is CPLEX-effective.) Moreover, the number of container operations in the sequence is a fixed number n, whereas in Meisel and Wichmann (2010)’s formulation (see Appendix A) the number of operations is upper bounded by n. (This may be another reason why our model is relatively competitive.) 5. Solution approach to large-scale problems CPLEX is not suitable for large-scale NP-hard problems. In this section, we devise a very time-effective heuristic and analyze its worst-case error bound. 5.1. A fast algorithm We design a constructive heuristic, called Internal-Reshuffle-Dense (IRD) algorithm. The intuitive idea is to maximize the number of internal-reshuffling operations. In particular, we update the digraph by reindexing all the chains in a decreasing

M. Liu et al. / Transportation Research Part E 81 (2015) 52–74

61

Fig. 8. Updated digraph of the example problem.

Fig. 9. IRD sequence of the example problem.

order of the number of reshuffles (see Fig. 8). Then we construct internal-reshuffling operations between two adjacent chains as many as possible (see Fig. 9). In the following, a job is called available if it appears at the head (i.e., first position) of a chain. Note that we assume reshuffles exist in the problem input, since otherwise the problem can be solved to optimum by Johnson’s rule (see Appendix B). Internal-Reshuffle-Dense (IRD) algorithm Step 1: Rearrange all chains in a decreasing order of the number of reshuffles. Let k ¼ 0. (Suppose there are m chains indexed from 1.) Step 2: k :¼ k þ 1. If k ¼ m, go to Step 5. Step 3: Process jobs in chain k until a BV job is available. If none, go to Step 2. Step 4: Process jobs in chain k þ 1 until a VB job is available. If none, go to Step 2. Process the VB job in chain k þ 1 and the BV job in chain k. Go to Step 3. Step 5: Process all the remaining jobs from top to bottom, chain by chain. As each stack has an equal number of reshuffles in both plans, chain rearrangement in Step 1 ensures that all VB jobs in chain k and an equal number of BV jobs in chain k  1 could construct internal-reshuffling operations, where k ¼ f2; . . . ; mg. In other words, all VB jobs in chain k are internally reshuffled. For example, jobs 5 and 1 in Fig. 9 are internally reshuffled. Remark 2. IRD algorithm takes Oðmaxfm log m; ngÞ time, which is polynomial in input n, where m 6 n. Let a denote the number of internal-reshuffling operations in a sequence (e.g., in Fig. 9, a ¼ 2). Use s0 to denote the time an internal-reshuffling takes (e.g., according to Table 4, s0 ¼ 110). We use s1 to indicate the setup time between an import container and an export container (no matter which goes first), and s2 the setup time between two import (or export)

62

M. Liu et al. / Transportation Research Part E 81 (2015) 52–74

containers (e.g., according to Table 4, s1 ¼ 10 and s2 ¼ 20). Clearly, s1 < s2 . Each job has a processing time p (e.g., according to Table 3, p ¼ 100). Recall that n denotes the number of containers in both plans. Theorem 1. The error bound of IRD algorithm is

ðn1aÞðs2 s1 Þ npþas0 þðn1aÞs1 .

Proof. Let C A denote the makespan of an IRD sequence, C OPT an optimal objective value. Clearly, C A 6 np þ as0 þ ðn  1  aÞs2 and C OPT P np þ as0 þ ðn  1  aÞs1 . Therefore the error bound is

C A  C OPT ðn  1  aÞðs2  s1 Þ 6 : np þ as0 þ ðn  1  aÞs1 C OPT h Corollary 1. The error bound of IRD algorithm is 1/10 with regard to the data settings in Meisel and Wichmann (2010). Proof. Let C A and C OPT respectively denote the objective value achieved by IRD algorithm and the optimal solution. By Theorem 1 and the data settings in Meisel and Wichmann (2010) (also see Tables 3 and 4), we know

C A  C OPT ðn  1  aÞðs2  s1 Þ na1 6 ¼ : np þ as0 þ ðn  1  aÞs1 11n  12a  1 C OPT na1 Let f ðn; aÞ ¼ 11n12 a1. By definitions of n (number of containers in both plans) and a (number of internal-reshuffling operations), we have n P 1 and a 6 bn=2c. Clearly, n P 2a. For fixed a, the partial derivative of f ðn; aÞ is

@f ðn; aÞ ðs2  s1 Þðas0 þ ap þ pÞ 10  a ¼ ¼ : 2 @n ððp þ s1 Þn þ ðas0  as1  s1 ÞÞ ð11n þ 12a þ 1Þ2 We discuss three cases of a. aÞ Case 1: 0 6 a < 10. In this case, we know @f ðn; > 0 and the function f ðn; aÞ attains the maximum as n approaches infinity, @n i.e.,

max f ðn; aÞ ¼ lim

na1 1 1 ¼ < :  12a  1 11 10

n!1 11n

n11 1 1 Case 2: a ¼ 10. In this case, f ðn; aÞ ¼ 11n121 ¼ 11 < 10 .

Case 3: a > 10. In this situation, we know

max f ðn; aÞ ¼ f ð2a; aÞ ¼

@f ðn;aÞ @n

< 0 and the function f ðn; aÞ attains the maximum when n ¼ 2a, i.e.,

a1 1 9 1 ¼  < : 10a  1 10 100a  10 10

This completes the proof. h 5.2. Comparison via example problem We use the example provided in Meisel and Wichmann (2010) to demonstrate how IRD algorithm works. Then, based on this example, we compare the results generated with IRD algorithm, Johnson’s rule (Goodchild and Daganzo, 2006), and GRASP heuristic (Meisel and Wichmann, 2010). In the example (see Fig. 3), there are five import containers, two export containers, four reshuffles, and two fixed containers. As fixed containers do not affect the decision process, we omit them in the concise representation (Fig. 5). The task is to convert the arrival plan to the departure plan. Below, we detail the solutions generated with different approaches.  A sequence generated with IRD algorithm. We illustrate the details of how IRD algorithm works by Fig. 10, or to say how the IRD sequence, 7–8–9–4–5–13–14–15–6–11–1–12–2–10–3, depicted by Fig. 9 corresponds to operating time. In Fig. 10, arrows denote the moving direction of the quay crane, and dotted box indicates the starting slot of a container operation. In total, there are 13 container operations, and the total operating time is 1460 s. Moreover, we also depict where internal-reshuffling operations and double-cycling occur. This sequence has two internal-reshuffling operations (i.e., 5–13 and 1–12) and double-cycling happens twice (i.e., 6–11 and 2–10). In-bay empty crane move occurs four times (i.e., 15–6, 11–1, 12–2 and 10–3).

M. Liu et al. / Transportation Research Part E 81 (2015) 52–74

63

Fig. 10. IRD sequence: 7–8–9–4–5–13–14–15–6–11–1–12–2–10–3.

 A sequence generated with Johnson’s rule. It is well known that Johnson’s rule can be applied to solve the QCDCP (Goodchild and Daganzo, 2006). For the given example, the Johnson-rule sequence is 7–8–9–4–13–5–14–6–15–1–11– 2–12–3–10. In total, there are 15 operations, and the time is 1670. In this sequence, double-cycling happens six times (i.e., 4–13, 5–14, 6–15, 1–11, 2–12, and 3–10), as expected, because the purpose of Johnson’s rule is to maximize the number of double cycles. In-bay empty crane move happens five times (i.e., 13–5, 14–6, 15–1, 11–2 and 12–3).

64

M. Liu et al. / Transportation Research Part E 81 (2015) 52–74

 A sequence generated with GRASP heuristic. As Meisel and Wichmann (2010) devise a GRASP heuristic to solve the problem. The GRASP sequence given in their work is 4–7–8–9–1–13–5–14–6–15–11–2–12–3–10. There are 13 container operations, and the objective value is 1450. In this sequence, internal reshuffling operation occurs twice (i.e., 1–13, 5–14), double-cycling happens three times (i.e., 6–15, 2–12 and 3–10). Besides, in-bay empty move occurs four times (i.e., 13–5, 14–6, 11–2 and 12–3).  A sequence generated with CPLEX. For the above example, an optimal solution can be obtained by CPLEX with either Meisel and Wichmann (2010)’s model or our GBIP model (see Section 4.2). An optimal solution is 7–8–9–13–4–5–14– 6–11–1–12–2–10–3–15. To save space, we do not depict the details of operations. In total, there are 13 container operations, in which internal-reshuffling operation occurs twice (i.e., 5–14 and 1–12). Double-cycling happens four times (i.e., 9–13, 6–11, 2–10 and 3–15). The total time consumed is 1430. In the optimal sequence, in-bay empty move occurs five times (i.e., 13–4, 14–6 11–1, 12–2 and 10–3). Table 5 sums up the results. The first column reports the objective function values. The second presents the optimality gap (compared with CPLEX solution). The following three columns illustrate the number of internal-reshuffling operations, the number of double-cycling and the number of in-bay empty moves, respectively. The last column illustrates time complexity of different methods. By comparison, we observe that (i) internal-reshuffling operations help to reduce service time by 16.8% with the sole application of the double cycling technique, (ii) IRD solution is better than Johnson-rule solution by 14.7%, (iii) IRD solution has an optimality gap 2.1%, and (iv) IRD algorithm’s performance is slightly weaker than that of GRASP by 0.7%. As for time complexity, (i) IRD algorithm is approximately equal to Johnson’s rule and (ii) IRD algorithm is significantly superior to GRASP heuristic. Besides, IRD algorithm generates the same number of internal-reshuffling operations as CPLEX does. Remark 3. Notice that Johnson’s rule generates an optimal solution to the problem without internal-reshuffling operations (i.e., QCDCP), whereas the CPLEX solution is an optimal solution to the problem with internal-reshuffling operations (i.e., QCDCP-IR). For this example, the internal-reshuffling method reduces the total service time by 16.8%, and it reduces the operations related to the four reshuffles by ð4  2  ð2 þ 2  2ÞÞ=ð4  2Þ ¼ 25%.

6. Computational experiments In this section, we conduct computational experiments for the following purposes: (i) to evaluate the quality of GBIP model presented in Section 4.2, compared with the start-of-the-art model and (ii) to assess the solution quality of IRD algorithm, compared with GRASP heuristic. IRD algorithm is coded in C++ language. GRASP heuristic is reimplemented according to the description in Meisel and Wichmann (2010) in C++ language. CPLEX 12.5 is used to solve the two integer programming models. All the experiments are conducted on a Macbook with 2.4 GHz and 8 GB RAM (with software Xcode 4.6.3). 6.1. Instance generation Experiments are conducted on a large set of test instances. 6.1.1. Bay sizes All the bay sizes description in Meisel and Wichmann (2010) are included in our experiments. Besides, to evaluate the efficiency of our GBIP model, an ‘‘extra small’’ bay size is designed, as CPLEX cannot generate optimal solutions for relatively large instances. To accommodate mega-ships, such as Maersk Triple-E class of mega-ship with TEU capacity of 18,270 (see Port Technology International, 2015), we also add an ‘‘extra large’’ bay size (see Table 6). 6.1.2. Workload scenarios Meisel and Wichmann (2010) generate three workload scenarios, i.e., high load, low import, and low export, which are distinguished by the ratio of import containers, export containers, reshuffles and fixed containers in the arrival and departure plans. In the example given in Fig. 3, there are five import containers, two export containers, four reshuffles (in either plan)

Table 5 Comparison between IRD, Johnson’s rule, GRASP and CPLEX.



Methods

Objective value (sec.)

Optimality gap (%)

#Internalreshuffling

# Doublecycling

#In-bay empty move

Time complexity

IRD Johnson’s rule GRASP CPLEX

1460 1670 1450 1430⁄

2.1 16.8 1.4 0

2 0 2 2

2 6 3 4

4 5 4 5

Oðmaxfm log m; ngÞ Oðm log mÞ Non-polynomial Non-polynomial

Optimal objective function value.

65

M. Liu et al. / Transportation Research Part E 81 (2015) 52–74 Table 6 Bay sizes for instance generation. Bay sizes

No. of stacks

No. of tiers

extra small small medium large extra large

5 10 15 20 25

5 10 15 20 25

Table 7 Workload scenarios for instance generation. Workload scenario

Import containers (%)

Export containers (%)

Reshuffles (%)

Fixed containers (%)

High load Low import Low export Low load

70 40 70 40

70 70 40 40

2–20 2–20 2–20 5–50

10 10 10 10

Table 8 Computational results of extra-small bay size. Workload scenario

RR (%)

Value

Time (sec.)

#Opt.

Value

Time (sec.)

#Opt.

Value

Time (sec.)

Value

Time (sec.)

High load

2 4 6 8 10 12 14 16 18 20

4450 4450 4450 4400 4450 4400 4490 4160 4130 4450

292 5 5 303 5 538 328 600 600 589

5 5 3 4 2 2 3 1 1 2

4450 4450 4450 4400 4450 4400 4410 4270 4180 4470

387 175 174 388 175 600 524 600 600 600

3 5 3 3 1 0 3 0 1 2

4620 4500 4525 4585 4510 4590 4420 4290 4200 4530

23 147 10 23 148 22 10 10 26 4

4685 4690 4690 4690 4690 4615 4625 4275 4265 4640

<0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01

4390 4390 4390 4330 4390 4330 4270 4090 4030 4390

Low import

2 4 6 8 10 12 14 16 18 20

3370 3370 3370 3370 3250 3320 3370 3370 3260 3150

32 22 10 15 142 138 36 71 110 226

5 5 5 5 5 5 5 5 5 5

3370 3370 3370 3370 3250 3320 3370 3370 3260 3150

47 33 38 39 312 202 70 185 333 490

5 5 5 5 5 5 5 5 4 3

3500 3470 3525 3460 3370 3345 3460 3430 3290 3200

5 84 5 86 5 30 5 5 86 5

3520 3525 3530 3530 3455 3475 3525 3525 3480 3245

<0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01

3290 3290 3290 3290 3170 3230 3290 3290 3170 3050

Low export

2 4 6 8 10 12 14 16 18 20

3330 3330 3330 3330 3330 3270 3220 3100 3330 3170

45 2 84 86 80 144 301 600 82 600

5 5 5 4 5 5 4 3 4 2

3330 3330 3330 3330 3330 3270 3220 3100 3330 3170

50 7 97 172 60 257 304 600 99 600

5 5 5 4 5 4 3 1 4 2

3400 3405 3425 3400 3410 3350 3310 3175 3465 3230

12 205 12 208 12 207 12 12 207 501

3480 3485 3485 3490 3490 3490 3415 3225 3490 3280

<0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01

3290 3290 3290 3290 3290 3230 3170 3050 3290 3110

Low load

5 10 15 20 25 30 35 40 45 50

2690 2590 2690 2630 2580 2580 2460 2400 2460 2470

6 52 5 16 51 119 85 65 58 71

5 5 5 5 5 5 5 5 5 5

2690 2590 2690 2630 2580 2580 2460 2400 2460 2470

4 52 4 32 115 159 163 278 133 115

5 5 5 5 5 5 5 5 5 5

2700 2720 3000 2970 2715 2800 2565 2575 2725 2605

3 53 182 3 53 1 11 62 206 1

2810 2710 2760 2710 2610 2645 2610 2460 2610 2565

<0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01

2630 2510 2630 2570 2510 2510 2390 2330 2390 2390

3460

181

4.25

3460

251

3.90

3494

12

3625

<0.01

3390

Average

GBIP model (CPLEX)

M&W model (CPLEX)

GRASP

IRD

GBIP’s improvement on Time = (251–181)/251 = 28%, improvement on #Opt. = (4.25–3.90)/3.90 = 9% Average Gaps of GRASP and IRD (%)

4.13

7.02

Lower bound

66

M. Liu et al. / Transportation Research Part E 81 (2015) 52–74

and two fixed containers. The capacity of a bay is 12 containers or slots (4 stacks  3 tiers). The ratio of import containers in the arrival plan is 5=12  42%, the ratio of export containers in the departure plan is 2=12  17%, the ratio of reshuffles is 4=12  33% in either plan. In their computational experiments, the percentage of fixed containers is set to 10% in all scenarios and the percentage of reshuffles varies in the range [0, 2%, . . ., 20%]. Note that by Theorem 2 in Appendix B, we show that an optimal solution to such a situation (with no reshuffles) can be obtained by Johnson’s rule. Thus, it is not necessary to test the cases with 0% reshuffles. Besides, to study the cases where reshuffles pose a high percentage, we extensively design a fourth scenario, called low load (see Table 7). In the fourth scenario, the percentage of fixed containers is set to 10% as in other three scenarios, and the percentage of reshuffles varies in the range [5%, 10%, . . ., 50%]. Thus, in each workload scenario, there are 10 values of reshuffle ratios or percentages. Table 7 summarizes the workload scenarios used in our computational study. In our experiments, there are in total 200 combinations (5 bay sizes  4 workload scenarios  10 reshuffle percentages). For each combination, five instances are generated by randomly assigning the different types of containers to slots in the plans. In total, 1000 benchmark instances are generated. In our tests, the data illustrated in Tables 1 and 2 are used, as in Meisel and Wichmann (2010). 6.2. Computational results To demonstrate that our GBIP model (see Section 4.2) is CPLEX-effective, we compare it with Meisel and Wichmann (2010)’s formulation, called M&W model, both solved with CPLEX. For some cases of extra small bay size, CPLEX cannot

Table 9 Computational results of small bay size. Workload scenario

RR (%)

GBIP model (CPLEX)

GRASP

IRD

Value

Time (sec.)

#Int.

Value

Time (sec.)

Lower bound

Gap (%)

Value

Time (sec.)

Gap (%)

High load

2 4 6 8 10 12 14 16 18 20

None None None 18,990 None 18,920 18,870 18,990 None 18,990

600 600 600 600 600 600 600 600 600 600

2 4 4 5 4 5 5 5 4 5

18,230 18,100 18,260 17,970 17,425 17,510 17,370 17,040 16,350 17,015

466 546 460 349 267 523 255 107 297 83

4.17 3.81 4.65 5.79 3.53 4.34 2.88 3.11 2.32 4.38

18,915 18,910 18,915 17,845 17,935 18,270 17,770 17,345 17,275 17,345

<0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01

8.27 8.62 8.64 5.40 6.69 9.07 5.33 5.06 8.17 6.61

17,470 17,410 17,410 16,930 16,810 16,750 16,870 16,510 15,970 16,270

Low import

2 4 6 8 10 12 14 16 18 20

14,250 14,190 14,270 14,250 14,170 14,270 14,200 14,130 14,270 14,270

600 600 600 600 600 600 600 600 600 600

5 5 5 5 5 5 5 5 5 5

13,645 13,425 13,035 13,745 12,740 13,125 13,140 13,345 12,890 12,285

139 320 128 135 362 439 560 409 140 382

5.09 3.54 2.03 5.57 1.65 4.53 3.27 4.31 4.19 3.87

14,050 13,765 13,645 14,200 13,200 13,410 13,580 13,625 13,270 12,625

<0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01

8.49 6.29 6.85 9.40 5.35 7.02 6.85 6.70 7.45 6.90

12,950 12,950 12,770 12,980 12,530 12,530 12,710 12,770 12,350 11,810

Low export

2 4 6 8 10 12 14 16 18 20

13,570 13,820 13,970 14,180 14,050 14,080 14,030 13,960 14,190 14,190

511 600 600 600 600 600 600 600 600 600

5 5 5 5 5 5 5 5 5 5

13,705 13,600 13,490 13,150 13,285 13,235 13,225 13,085 12,170 12,260

199 96 415 167 167 209 479 129 434 104

3.39 3.01 4.00 4.26 3.88 3.51 5.71 3.78 1.97 4.16

14,190 14,190 13,790 13,190 13,190 13,190 13,690 13,190 12,690 12,690

<0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01

7.18 7.58 6.49 4.77 3.29 3.29 9.78 4.77 6.37 8.00

13,240 13,190 12,950 12,590 12,770 12,770 12,470 12,590 11,930 11,750

Low load

5 10 15 20 25 30 35 40 45 50

10,890 11,240 11,360 11,350 11,160 11,390 11,380 11,390 11,390 11,390

600 600 600 600 600 600 600 600 600 600

5 5 5 5 5 5 5 5 5 5

11,530 10,170 11,045 10,490 10,965 10,610 9815 9495 9775 9255

245 268 499 269 439 265 479 151 417 323

8.41 2.16 11.00 4.58 8.71 12.44 5.35 5.32 8.64 9.35

11,390 10,600 10,390 10,890 10,390 9890 10,090 9790 9750 8950

<0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01 <0.01

7.86 6.53 5.70 8.79 3.80 6.46 8.61 8.90 9.18 6.67

10,560 9950 9830 10,010 10,010 9290 9290 8990 8930 8390

13742.5

395

5.91

13932.5

<0.01

7.75

12,930

Average

Note. Value ‘‘None’’ means CPLEX does not deliver integer solutions for some instances in one combination.

67

M. Liu et al. / Transportation Research Part E 81 (2015) 52–74

optimally solve M&W model in 3600 s. Thus, we set a time limit of 600 s in CPLEX for both models, and count the number of optimal solutions obtained in the tests. Lower bounds can be obtained by the procedure provided by Meisel and Wichmann (2010), who demonstrate the bounds generated with their procedure are better than those generated with CPLEX in one hour in most of the cases. Therefore, we use this procedure to calculate lower bounds. Table 8 shows the computational results of extra small bay size in all four scenarios. Every row in this table reports average results for a combination of five instances. Column ‘‘Workload scenario’’ indicates the four workload scenarios. Column ‘‘RR (%)’’ gives the reshuffle ratios or percentages. The third to fifth columns report the average objective function values obtained by solving GBIP model with CPLEX, and the average running times, and the number of optimal values among five instances, respectively. The sixth to eighth columns report those of M&W model. The ninth to eleventh (twelfth to fourteenth) columns represent the average objective function values obtained with GRASP heuristic (IRD algorithm) and its average running times, respectively. The last column reports the average lower bounds delivered by the lower-bounding procedure (Meisel and Wichmann, 2010). In the experiments, the units of objective function values and running time are seconds (sec. for short). Results in Table 8 demonstrate that on average using GBIP model CPLEX can optimally solve 4.25 instances (of five instances in each combination) in the time limit, whereas this value of M&W model is 3.90. Thus, in terms of solution quality, compared with M&W model, GBIP model can obtain 9% more optimal solutions in the time limit. In terms of running time, results show that GBIP model has an improvement about 28%, compared with M&W model. In this table, the superiority of GBIP model is clearly shown. The ‘‘Average Gap (%)’’ in the last row reports the relative error gaps of IRD and GRASP for all the

Table 10 Computational results of medium bay size. Workload scenario

RR (%)

GBIP model (CPLEX)

GRASP

IRD

Value

Time (sec.)

#Int.

Value

Time (sec.)

Lower bound

Gap (%)

Value

Time (sec.)

Gap (%)

High load

2 4 6 8 10 12 14 16 18 20

None None None None None None None None None 42,890

600 600 600 600 600 600 600 600 600 600

0 0 3 3 2 3 2 3 3 5

39,675 41,175 40,110 40,190 39,745 39,780 38,370 38,495 37,260 38,130

569 334 584 752 635 345 475 626 567 323

1.29 4.96 2.87 5.51 3.69 5.77 2.84 5.38 3.70 5.77

41,300 42,815 42,745 41,370 41,370 38,945 39,500 38,455 38,010 38,005

<1 <1 <1 <1 <1 <1 <1 <1 <1 1

5.44 9.14 9.63 8.61 7.93 3.55 5.87 5.27 5.79 5.42

39,170 39,230 38,990 38,090 38,330 37,610 37,310 36,530 35,930 36,050

Low import

2 4 6 8 10 12 14 16 18 20

None None None None None None None None None None

600 600 600 600 600 600 600 600 600 600

0 1 0 2 2 1 4 1 1 3

31,080 29,775 30,265 29,890 29,870 29,910 29,290 28,960 28,200 28,605

455 79 398 519 646 262 346 165 246 486

5.53 2.99 3.19 5.14 4.62 6.10 3.68 4.06 4.02 4.36

30,050 30,000 32,135 29,935 31,105 30,210 30,105 30,210 29,000 28,650

<1 <1 <1 <1 <1 <1 <1 <1 <1 1

2.04 3.77 9.56 5.29 8.95 7.17 6.57 8.55 6.97 4.52

29,450 28,910 29,330 28,430 28,550 28,190 28,250 27,830 27,110 27,410

Low export

2 4 6 8 10 12 14 16 18 20

None None None None None None 32,060 32,090 32,090 None

600 600 600 600 600 600 600 600 600 600

1 2 2 3 2 4 5 5 5 3

30,580 30,210 29,595 30,070 29,040 30,570 29,480 29,405 28,295 27,990

419 643 537 61 441 414 308 611 219 453

3.84 3.21 1.53 5.32 2.15 7.08 5.93 5.21 4.14 2.34

32,015 32,080 31,090 30,085 31,090 30,090 28,660 30,375 29,825 29,345

<1 <1 <1 <1 <1 <1 <1 <1 <1 <1

8.71 9.60 6.66 5.38 9.36 5.39 2.98 8.68 9.77 7.29

29,450 29,270 29,150 28,550 28,430 28,550 27,830 27,950 27,170 27,350

Low load

5 10 15 20 25 30 35 40 45 50

None None None None None None 25,730 None 25,730 25,730

600 600 600 600 600 600 600 600 600 600

1 2 2 2 4 4 5 4 5 5

26,830 23,825 24,530 23,850 24,435 22,380 21,150 22,105 19,575 19,885

299 108 263 319 667 167 609 695 311 313

14.71 4.82 7.92 9.86 13.49 9.76 5.28 12.04 2.33 6.97

25,585 24,730 23,590 23,230 23,230 22,230 21,910 20,575 20,445 19,980

<1 <1 <1 <1 <1 <1 <1 <1 <1 <1

9.38 8.80 3.78 7.00 7.90 9.02 9.06 4.28 6.87 7.48

23,390 22,730 22,730 21,710 21,530 20,390 20,090 19,730 19,130 18,590

29,780

441

4.13

30,640

<1

6.46

28,880

Average

Note. Value ‘‘None’’ means CPLEX does not deliver integer solutions for some instances in one combination.

68

M. Liu et al. / Transportation Research Part E 81 (2015) 52–74

combinations, where a relative error gap of each combination is calculated by ðValue  Lower boundÞ=Lower bound  100%. In terms of solution quality, on average, IRD algorithm has a relative error bound of 7.02%, slightly weaker than that of GRASP by 2.89%. In terms of running time, IRD algorithm clearly outperforms GRASP, with a running time far less than one second, whereas GRASP consumes 181 s on average. Table 9 presents the results of small bay size. We only report the CPLEX solutions solving with GBIP model, as M&W model is less CPLEX-effective, shown previously. The purpose is to report the best solutions of CPLEX in 600 s for the instances. Column ‘‘#Int.’’ reports the number of integer (feasible) solutions obtained for each combination of five instances, as for some instances CPLEX cannot deliver integer solutions. In Column ‘‘Value’’, we present the average objective function values where integer solutions are obtained for all five instances, as CPLEX cannot generate integer solutions for some instances. Column ‘‘Gap (%)’’ gives the relative error gap for each combination. Results show that CPLEX can obtain feasible solutions for most instances of small bay size, however these objective values are not competitive, compared with the values generated with IRD algorithm. Generally, the GRASP heuristic outperforms the IRD algorithm in terms of solution quality, as IRD algorithm has a relative error gap 7.75% on average, slightly weaker than that of GRASP by 1.84%. Notice that for some combinations, the relative error gap of GRASP is larger than 10%, whereas the IRD algorithm guarantees a relative error bound less than 10%. Thus, the IRD algorithm is relatively robust. In terms of running time, IRD algorithm clearly outperforms GRASP, with running time less than one second, whereas GRASP consumes 395 s. On average, in terms of solution quality, IRD outperforms GRASP in the low load scenario. Table 10 reports the computational results of the instances of medium bay size. As CPLEX has not totally lost its power for small bay sized instances, we also report the solutions generated with CPLEX (with GBIP model). Results show that CPLEX

Table 11 Computational results of large bay size. Workload scenario

RR (%)

GRASP

IRD

Value

Time (sec.)

Lower bound

Gap (%)

Value

Time (sec.)

Gap (%)

High load

2 4 6 8 10 12 14 16 18 20

70,635 72,140 70,540 68,965 68,855 67,550 68,505 68,400 66,825 65,965

20 220 613 587 404 274 295 441 645 360

1.74 5.54 3.93 2.06 3.37 2.71 4.73 4.76 4.86 4.49

72,240 71,625 69,145 69,085 71,390 70,390 71,390 68,390 66,390 66,890

<1 <1 <1 <1 <1 <1 <1 <1 1 1

4.05 4.79 1.88 2.24 7.18 7.02 9.14 4.75 4.17 5.96

69,430 68,350 67,870 67,570 66,610 65,770 65,410 65,290 63,730 63,130

Low import

2 4 6 8 10 12 14 16 18 20

53,900 54,810 52,945 51,790 51,235 50,675 51,530 51,300 50,285 50,625

334 234 273 112 402 150 216 686 284 481

3.75 5.51 2.51 1.21 1.92 1.41 4.63 4.04 4.91 4.32

54,200 54,200 54,200 53,275 54,830 54,350 52,350 53,350 50,440 50,350

<1 <1 <1 <1 <1 <1 <1 <1 <1 <1

4.33 4.33 4.94 4.11 9.07 8.77 6.29 8.19 5.24 3.75

51,950 51,950 51,650 51,170 50,270 49,970 49,250 49,310 47,930 48,530

Low export

2 4 6 8 10 12 14 16 18 20

54,885 51,930 52,850 53,415 53,480 51,525 50,500 50,705 50,105 48,710

110 428 223 296 651 777 171 574 217 171

4.68 1.37 3.53 5.88 5.50 4.75 2.79 5.53 3.25 2.92

56,115 55,190 53,765 53,205 53,630 52,625 51,215 50,615 51,190 51,965

<1 <1 <1 <1 <1 <1 <1 <1 <1 <1

7.03 7.73 5.32 5.46 5.80 6.98 4.24 5.34 5.48 9.79

52,430 51,230 51,050 50,450 50,690 49,190 49,130 48,050 48,530 47,330

Low load

5 10 15 20 25 30 35 40 45 50

44,390 43,000 43,720 42,350 38,715 38,580 39,405 38,045 35,435 34,445

515 311 337 639 203 454 268 332 450 441

8.51 7.31 11.45 8.95 2.61 6.13 11.16 10.50 3.46 4.60

42,520 43,970 41,970 41,540 41,040 38,430 38,415 36,470 35,400 35,400

<1 <1 <1 <1 <1 <1 <1 <1 <1 <1

3.94 9.73 6.98 6.87 8.77 5.72 8.36 5.93 3.36 7.50

40,910 40,070 39,230 38,870 37,730 36,350 35,450 34,430 34,250 32,930

52,540

231

3.17

53,820

<1

5.77

51,180

Average

69

M. Liu et al. / Transportation Research Part E 81 (2015) 52–74 Table 12 Computational results of extra large bay size. Workload scenario

RR (%)

GRASP

IRD

Value

Time (sec.)

Gap (%)

Lower bound

Value

Time (sec.)

Gap (%)

High load

2 4 6 8 10 12 14 16 18 20

112,040 111,140 109,840 110,010 108,680 106,970 106,175 104,705 104,605 98,260

96 504 620 540 642 448 372 309 276 197

3.10 3.18 2.66 4.23 3.61 4.13 3.78 3.19 4.64 1.21

115,200 114,700 114,700 114,340 109,340 109,270 109,770 110,200 107,270 106,050

<1 <1 <1 <1 <1 1 1 1 2 2

6.01 6.49 7.21 8.33 4.24 6.37 7.29 8.60 7.30 9.23

108,670 107,710 106,990 105,550 104,890 102,730 102,310 101,470 99,970 97,090

Low import

2 4 6 8 10 12 14 16 18 20

85,395 84,010 85,515 83,885 81,405 78,620 79,585 77,165 77,625 77,680

98 345 480 369 317 514 244 496 317 517

4.28 3.50 6.06 6.57 3.82 1.98 3.48 2.16 3.27 5.53

89,615 86,400 87,470 81,690 82,615 83,540 79,995 80,190 80,615 79,840

<1 <1 <1 <1 <1 <1 <1 <1 1 1

9.43 6.44 8.48 3.79 5.36 8.37 4.01 6.17 7.24 8.46

81,890 81,170 80,630 78,710 78,410 77,090 76,910 75,530 75,170 73,610

Low export

2 4 6 8 10 12 14 16 18 20

86,645 83,520 83,510 81,685 83,565 82,380 79,990 79,755 79,655 75,835

420 372 692 327 536 427 413 418 426 548

6.43 3.05 4.04 3.62 5.29 5.55 3.60 4.11 4.71 2.44

89,410 86,490 86,490 82,940 85,990 82,940 82,715 80,990 79,415 79,340

<1 <1 <1 <1 <1 1 1 1 1 1

9.83 6.71 7.75 5.21 8.34 6.27 7.13 5.72 4.40 7.17

81,410 81,050 80,270 78,830 79,370 78,050 77,210 76,610 76,070 74,030

Low load

5 10 15 20 25 30 35 40 45 50

69,945 66,135 65,990 61,670 60,865 58,430 61,260 57,515 53,735 54,325

29 380 530 297 356 610 634 382 123 358

8.36 5.80 10.00 4.05 6.35 3.29 11.85 7.61 5.01 8.07

69,545 65,540 63,905 61,970 60,040 59,540 56,980 58,760 54,980 54,980

<1 <1 <1 <1 <1 <1 <1 <1 <1 1

7.74 4.85 6.53 4.56 4.91 5.25 4.04 9.93 7.45 9.37

64,550 62,510 59,990 59,270 57,230 56,570 54,770 53,450 51,170 50,270

83,182

227

5.58

85,090

<1

7.69

79,470

Average

cannot deliver integer solutions for most of the instances. The average relative error gap of IRD is 6.46%, slightly weaker than that of GRASP by 2.33%. The maximum relative error gap of GRASP is 14.71%, whereas that value of IRD is 9.56%. Therefore, IRD algorithm is relatively robust. In terms of running time, IRD consumes less than one second, whereas GRASP takes 441 s on average. Thus, IRD saves a lot of computational time. On average, in terms of solution quality, IRD outperforms GRASP in the low load scenario. The computational results of the instances of large bay size is given in Table 11. As shown previously that CPLEX cannot generate integer solutions for most of the instances of medium bay size in the time limit, we only reports solutions generated with GRASP and IRD. On average, the relative error gap IRD is 5.77%, slightly weaker than that of GRASP by 2.60%. In Table 12, the instances of extra large bay size are tested. On average, IRD has a relative error gap 7.69%, slightly weaker than that of GRASP by 2.11%. On average, in terms of solution quality, IRD outperforms GRASP in the low load scenario. Summing up, (i) our GBIP model is CPLEX-effective compared with Meisel and Wichmann (2010)’s model, (ii) IRD algorithm, taking about one second to solve problems of practical size, is very efficient in terms of running time, (iii) IRD algorithm is relative robust with a relative error gap less than 10%, (iv) on average, in terms of solution quality, IRD outperforms GRASP in the scenario with high reshuffle percentages, and (v) with the increase of problem sizes, the difference between the two average relative error gaps decreases. Our experiment results also demonstrate that although the solution quality of IRD algorithm is slightly weaker than that of GRASP heuristic by about 2%, IRD algorithm can save a lot of computational time. Moreover, as IRD is a constructive heuristic, it is very easy to implement.

70

M. Liu et al. / Transportation Research Part E 81 (2015) 52–74

7. Conclusion In this paper, we revisit the quay crane double cycling problem with internal-reshuffling operations. For the problem, we propose some observations to map container operations to containers in the arrival and departure plans. Then we present a new integer programming model. Compared with Meisel and Wichmann (2010)’s formulation, experimental results demonstrate our formulation is CPLEX-effective. To efficiently solve real-world problem instances, we devise a constructive heuristic approach with worst-case error guarantee. Experimental results demonstrate that our algorithm is very fast and relative robust than the existing method. As the internal-reshuffling method can be used to complement the classic double-cycling technique to further improve QC efficiency, our algorithm is very attractive because it is low-cost and very easy to implement. Future research directions may include the general problems where the following assumptions are overcome: (i) every container to be reshuffled in the arrival plan can be positioned at any slot reserved for reshuffles in the departure plan, (ii) the number of containers under reshuffling in one stack is the same in both plans, and (iii) the hatch-covered problems are not considered. Moreover, (i) exact solution approaches, such as branch-and-price algorithm, are needed for the QCDCP-IR and its generalization, and (ii) aiming at looking for a double-cycling operation while pursuing internal reshuffling, the possibility of inserting a look-ahead step within IRD should be further explored. Acknowledgement The authors would like to acknowledge the constructive comments by the referees and the editor. The authors thank Prof. Chung-Yee Lee for valuable suggestions. The work described in this paper was supported by the National Natural Science Foundation of China (Grant Nos. 71101106, 71428002, 71272045, and 71072026). Appendix A We restate the integer programming formulation proposed by Meisel and Wichmann (2010). Moreover, we correct some mistakes, i.e., constraints (16)–(18) are corrected. Meisel and Wichmann (2010)’s model with corrections In Meisel and Wichmann (2010)’s formulation, ði; jÞ denotes the container occupying the position in i-th stack and j-th tier in a plan (see Fig. 3). Note that such a definition of ði; jÞ differs from ours. Clearly, the container in slot ði; j þ 1Þ is laid on top of the container in slot ði; jÞ. Parameters: m: number of stacks in the bay; n: number of tiers in the bay; AI : set of slots that hold an import container in the arrival plan; AR : set of slots that hold a reshuffle in the arrival plan; S A: A ¼ AI AR , i.e., set of slots that hold an import container or a reshuffle in the arrival plan; E D : set of slots that hold an export container in the departure plan; DR : set of slots that hold a reshuffle in the departure plan; S D: D ¼ DE DR , i.e., set of slots that hold an export container or a reshuffle in the departure plan; T: set of container operation types, T ¼ fVY; YV; VB; BV; VVg; t d: processing time of a container operation of type t 2 T; tu d : time needed for empty crane spreader movement in-between a container operation of type t and a container operation of type u, where t; u 2 T; l: l ¼ jAj þ jDj, i.e., an upper bound on the number of container operations; K; K: K ¼ f1; 2; . . . ; lg and K ¼ f1; 2; . . . ; l  1g, i.e., index sets of container operations.

Variables: xVijk : equal 1 if the container at bay slot ði; jÞ is picked in move k 2 K; 0 otherwise; xYk :

equal to 1 if a container is picked from the yard in move k; 0 otherwise;

xBk :

equal to 1 if a container is picked from the buffer in move k; 0 otherwise;

yVijk :

equal to 1 if a container is dropped at bay slot ði; jÞ in move k; 0 otherwise;

yYk :

equal to 1 if a container is dropped at the yard in move k; 0 otherwise;

yBk :

equal to 1 if a container is dropped at the buffer in move k; 0 otherwise;

stk :

equal to 1 if container operation k is of type t 2 T; 0 otherwise;

M. Liu et al. / Transportation Research Part E 81 (2015) 52–74

71

etu k :

equal to 1 if container operation k is of type t 2 T and operation k þ 1 is of type u 2 T, i.e., stk ¼ sukþ1 ¼ 1; 0 otherwise;

bk :

number of externally buffered reshuffles at the end of operation k.

ðM&W modelÞ

min Z ¼

XX t X X tu d  stk þ d  etu k k2K t2T

X xVijk ¼ 1;

s:t:

ð13Þ

k2K t;u2T

8ði; jÞ 2 A:

ð14Þ

k2K

X yVijk ¼ 1;

8ði; jÞ 2 D:

ð15Þ

k2K

X X xVi;jþ1;k  k þ 1 6 xVijk  k; k2K

8ði; jÞ; ði; j þ 1Þ 2 A:

ð16Þ

8ði; jÞ; ði; j þ 1Þ 2 D:

ð17Þ

k2K

X X yVijk  k þ 1 6 yVi;jþ1;k  k; k2K

k2K

X X xVijk  k þ 1 6 yVijk  k; k2K

8ði; jÞ 2 A

\

D:

ð18Þ

k2K

X

stk 6 1; 8k 2 K:

ð19Þ

t2T

sVY k ¼

X

xVijk ;

8k 2 K:

ð20Þ

ði;jÞ2AI Y sVY 8k 2 K: k ¼ yk ;

ð21Þ

Y sYV 8k 2 K: k ¼ xk ;

ð22Þ

X

sYV k ¼

yVijk ;

8k 2 K:

ð23Þ

xVijk ;

ð24Þ

ði;jÞ2DE

X

VB sVV k þ sk ¼

ði;jÞ2A

8k 2 K:

R

B sVB 8k 2 K: k ¼ yk ;

ð25Þ

B sBV 8k 2 K: k ¼ xk ;

ð26Þ

X

BV sVV k þ sk ¼

ði;jÞ2D

X

X

t2T

t2T

stk P

yVijk ;

8k 2 K:

stkþ1 ; 8k 2 K:

t u etu k P sk þ skþ1  1;

ð28Þ

t; u 2 T; 8k 2 K:

b0 ¼ 0:

ð29Þ ð30Þ

bk ¼ bk1  xBk þ yBk ; bk 2 Z þ ;

ð27Þ

R

8k 2 K:

8k 2 K:

xVijk ; xYk ; xBk ; stk ; etu k 2 f0; 1g;

ð31Þ ð32Þ

8ði; jÞ 2 A; k 2 K:

ð33Þ

72

M. Liu et al. / Transportation Research Part E 81 (2015) 52–74

yVijk ; yYk ; yBk 2 f0; 1g;

8ði; jÞ 2 D; k 2 K:

ð34Þ

The objective function (13) is to minimize the service time of the bay, which is the sum of total time needed for container operations and total time needed for empty spreader move. Constraints (14) ensure that each import container and each reshuffle in the arrival plan is picked up once by the QC. Constraints (15) make sure that exactly one container is placed at every export and every slot to be hold by a reshuffle in the departure plan. Constraints (16) guarantee that the pick-up of slot ði; j þ 1Þ precedes the pick-up of slot ði; jÞ in the arrival plan. (We fixed this set of constraints in Meisel and Wichmann, 2010.) Constraints (17) ensure that the drop-off of slot ði; jÞ precedes the drop-off of slot ði; jÞ in the departure plan. Constraints (18) ensure that a container is removed from a slot before a new container is loaded to that place. (We fixed this set of constraints in Meisel and Wichmann, 2010) Constraints (19) make sure at most one container is moved in operation k 2 K. Constraints (20) and (21) guarantee that one import container (or one VY move) corresponds to one pick-up and one drop-off operations. Constraints (22) and (23) provide the relation between an export container and its pick-up and drop-off operations. Similarly, constraints (24)–(27) ensure the relation between one reshuffle and its pick-up and drop-off operations. Constraints (28) enforce a consecutive sequence of non-empty container operations in case that less than l moves are contained in a solution. Constraints (29) establish the relation between one empty spreader move and the related two container operations. Constraints (30) and (32) guarantee the number of reshuffles in the dock buffer is not less than zero. Domains of the decision variables are defined in (33) and (34). Remark 4. This formulation is a natural representation as the problem appears.

Appendix B In this part, we identify a special case for the QCDCP-IR which can be solved in polynomial time. The potential benefit may be that if an instance is recognized as such a case, then it is can be easily solved. B.1. The non-reshuffle QCDCP-IR We consider the case where no reshuffles exist. That is, the problem contains only import containers and export containers, and thus no internal-reshuffling operations could ever happen. We denote this special case by the nr-QCDCP-IR. Note that the nr-QCDCP-IR is different from the QCDCP, because in the QCDCP container’s operation time and crane’s empty move time (from the dock to the ship, or in a reverse direction) are identical, whereas the nr-QCDCP-IR involves different kinds of empty move times. Thus, the nr-QCDCP-IR is not trivial. For simplicity, let s1 denote the setup time between an import container and an export container (no matter which goes first), and s2 the setup time between two import (or export) containers. Clearly, s1 < s2 . Lemma 1. The nr-QCDCP-IR is to maximize the number of s1 .

Proof. This is because the number of setup times is a constant n  1, where n denotes the number of containers. The more time s1 appears in the sequence, the smaller the makespan is. h B.2. Two machine flow shop Following Goodchild and Daganzo (2006)’s work by extending their methodology, we regard each stack as a job (with two sequential operations) for the two machine flow shop problem. The first (second) operation of a specific job is constructed by concatenating the import (export) containers from top to bottom (bottom to top) in the arrival (departure) plan. Clearly, the processing time of its first (second) operation is the number of import (export) containers. See Fig. 11a for illustration. The flow shop problem allows preemption at integral time points, since each container is represented by a processing request of length one. Moreover, only active schedules are considered, i.e., there is no unnecessary inserted idleness in the schedule. Remark 5. All precedence constraints have been captured within flow shop job definition. We next map a flow shop schedule to a single machine sequence. Given a flow shop schedule, we can obtain a job sequence for single machine by the following procedure (e.g., see Fig. 11b). Mapping Procedure Step 1: Set time t ¼ 0. Step 2: Choose and add to the targeted sequence, from the schedule of flow shop, the job segment which starts at time t on machine 1 (if any), and then the job segment which starts at time t on machine 2 (if any). If no further processing exists, stop; otherwise, set t :¼ t þ 1 and go to Step 2.

M. Liu et al. / Transportation Research Part E 81 (2015) 52–74

73

Fig. 11. Two machine flow shop.

Lemma 2. The mapped single machine sequence is feasible. Proof. This is because there is no violation of precedence relations. h Lemma 3. In the single machine sequence, s1 occurs 2k  1 times, where k denotes the number of time units when both machines are simultaneously busy. Proof. Observe that s1 (in the mapped single machine sequence) occurs when the processing crosses machines in the flow shop schedule. For a given flow shop schedule with k common busy time units, assuming there is a virtual ‘‘line’’ between the two machines, then the processing of jobs crosses this line 2k  1 times. h Lemma 4. The nr-QCDCP-IR aims to maximize the number of k. Proof. By Lemma 1. h Lemma 5. The nr-QCDCP-IR is equivalent to the two machine flow shop makespan minimization problem. Proof. Observe that two machine flow shop makespan minimization problem is equivalent to maximize the number of time units when both machines are simultaneously busy. It is well known that preemption has no advantage for this flow shop problem. By Lemma 4, the proposition follows. h Recall that Johnson’s rule optimally solves two machine flow shop makespan minimization problem (Pinedo, 2012). Let uðiÞ (lðiÞ) denote the processing time of the first (second) operation of job J i . Johnson’s rule (i) If uðiÞ < lðiÞ, assign job J i to group A; otherwise, assign it to group B. (ii) Arrange jobs in group A in non-decreasing order of uðiÞ, and jobs in group B in non-increasing order of lðiÞ. (iii) Construct the final sequence in which all jobs in group A are followed by all jobs in group B.

Theorem 2. The nr-QCDCP-IR can be optimally solved by Johnson’s rule. Proof. By Lemma 5. h References Bell, M.G.H., Liu, X., Angeloudis, P., Fonzone, A., Hosseinloo, S.H., 2011. A frequency-based maritime container assignment model. Transp. Res. Part B 45 (8), 1152–1161. Bell, M.G.H., Liu, X., Rioult, J., Angeloudis, P., 2013. A cost-based maritime container assignment model. Transp. Res. Part B 58, 58–70. Bierwirth, C., Meisel, F., 2009. A fast heuristic for quay crane scheduling with interference constraints. J. Sched. 12 (4), 345–360. Bierwirth, C., Meisel, F., 2010. A survey of berth allocation and quay crane scheduling problems in container terminals. Eur. J. Oper. Res. 202 (3), 615–627.

74

M. Liu et al. / Transportation Research Part E 81 (2015) 52–74

Bierwirth, C., Meisel, F., 2015. A follow-up survey of berth allocation and quay crane scheduling problems in container terminals. Eur. J. Oper. Res. http:// dx.doi.org/10.1016/j.ejor.2014.12.030. Carlo, H.J., Vis, I.F.A., Roodbergen, K.J., 2013. Seaside operations in container terminals: literature overview, trends, and research directions. Flex. Serv. Manuf. J. http://dx.doi.org/10.1007/s10696-013-9178-3. Choo, S., Klabjan, D., Simchi-Levi, D., 2010. Multiship crane sequencing with yard congestion constraints. Transp. Sci. 44 (1), 98–115. Crainic, T.G., Kim, K.H., 2005. Intermodal transportation. In: Barnhart, C., Laporte, G. (Eds.), Transportation Handbooks in Operations Research and Management Science. North-Holland, Amsterdam, Netherlands. Daganzo, C.F., 1989. Crane productivity and ship delay in ports. Transp. Res. Rec. 1251, 1–9. Drewry Container Insight, 2014. . Goodchild, A.V., 2005. Crane Double Cycling in Container Ports: Algorithms, Evaluation, and Planning. Ph.D. thesis, Department of Civil and Environmental Engineering, University of California at Berkeley. Goodchild, A.V., Daganzo, C.F., 2006. Double-cycling strategies for container ships and their effect on ship loading and unloading operations. Transp. Sci. 40 (4), 473–483. Goodchild, A.V., Daganzo, C.F., 2007. Crane double cycling in container ports: planning methods and evaluation. Transp. Res. Part B: Methodol. 41 (8), 875– 891. Goodchild, A.V., McCall, J.G., Zumerchik, J., Lanigan Sr., J., 2011. Reducing train turn times with doubly cycling in new terminal designs. Transp. Res. Rec.: J. Transp. Res. Board 2238, 8–14. Kim, K.H., Park, Y.-M., 2004. A crane scheduling method for port container terminals. Eur. J. Oper. Res. 156 (3), 752–768. Lee, C-Y., Liu, M., Chu, C.B., 2014. Optimal algorithm for general quay crane double-cycling problem. Transp. Sci. http://dx.doi.org/10.1287/trsc.2014.0563. Lim, A., Rodrigues, B., Xu, Z., 2007. A m-parallel crane scheduling problem with a non-crossing constraint. Naval Res. Logist. 54 (2), 115–127. Liu, M., Zheng, F.F., Li, J.F., 2014. Scheduling small number of quay cranes with non-interference constraint. Optim. Lett. http://dx.doi.org/10.1007/s11590014-0756-4. Meisel, F., Wichmann, M., 2010. Container sequencing for quay cranes with internal-reshuffles. OR Spectrum 32 (3), 569–591. Meng, Q., Wang, S.A., 2011. Liner shipping service network design with empty container repositioning. Transp. Res. Part E: Logist. Transp. Rev. 47 (5), 695– 708. Meng, Q., Wang, S.A., Andersson, H., Thun, K., 2014. Containership routing and scheduling in liner shipping: overview and future research directions. Transp. Sci. 48 (2), 265–280. Pinedo, M., 2012. Scheduling: Theory, Algorithms, and Systems, fourth ed. Springer. Port Technology International, 2015. Friday Focus: Maersk Triple-E Infographic. . Stahlbock, R., Voß, S., 2008. Operations research at container terminals: a literature update. OR Spectrum 30 (1), 1–52. Wang, S.A., 2014. A novel hybrid-link-based container routing model. Transp. Res. Part E: Logist. Transp. Rev. 61, 165–175. Wang, S.A., Liu, Z., Bell, M.G.H., 2015. Profit-based maritime container assignment models for liner shipping networks. Transp. Res. Part B 72, 59–76. Wang, S.A., Meng, Q., 2012. Sailing speed optimization for container ships in a liner shipping network. Transp. Res. Part E 48 (3), 701–714. Wang, S.A., Meng, Q., Liu, Z., 2013. Bunker consumption optimization methods in shipping: a critical review and extensions. Transp. Res. Part E 53, 49–62. Wang, H., Wang, S.A., Meng, Q., 2014. Simultaneous optimization of schedule coordination and cargo allocation for liner container shipping networks. Transp. Res. Part E: Logist. Transp. Rev. 70, 261–273. Zhang, H.P., Kim, K.H., 2009. Maximizing the number of dual-cycle operations of quay cranes in container terminals. Comput. Ind. Eng. 56 (3), 979–992.